Tool Reviews

AI Coding Agent Pricing Compared (2026)

Compare 2026 AI coding agent plans, included usage, overages, and the limits that matter.

  • #AI coding agents
  • #pricing
  • #Claude Code
  • #Codex

AI coding-agent pricing is no longer a simple contest between monthly subscription fees. A $20 plan may include a variable allowance, a fixed credit pool, or access that stops at a rolling limit. Another plan may cost more but expose clearer metering. The practical comparison is therefore: seat price, included agent usage, how that usage is measured, and what happens when it runs out.

This guide compares individual plans from Claude Code, OpenAI Codex, GitHub Copilot, and Cursor, plus Google’s organization-focused Gemini Code Assist licenses. Prices, credits, and product specifications below are as of 2026-07-20. Vendors can change plans, regional prices, taxes, model multipliers, and usage policies, so verify the linked official pages before buying.

Quick comparison

Product and planList price in USDIncluded agent capacityWhat happens at the limit
Claude Pro$20/month or $200/yearClaude and Claude Code share the plan’s usage limitsWait for reset, upgrade, or deliberately enable additional usage credits/API billing
Claude Max 5x$100/monthFive times Pro capacity per session; weekly limits also applyWait, buy usage credits, or move to Max 20x
Claude Max 20x$200/monthTwenty times Pro capacity per session; weekly limits also applyWait or enable additional paid usage
ChatGPT Plus with Codex$20/monthIncluded Codex allowance; consumption varies by input, cached input, output, model, and taskBuy credits after the included allowance, if available for the account, or wait for reset
ChatGPT Pro 5x with Codex$100/monthFive times Plus usage; individual model allowances can differAdd credits, change model, or wait for reset
ChatGPT Pro 20x with Codex$200/monthTwenty times Plus usage; individual model allowances can differAdd credits, change model, or wait for reset
GitHub Copilot Free$0Limited chat and agent usage; 2,000 completions per monthWait for reset or upgrade
GitHub Copilot Pro$10/month$15 in total monthly GitHub AI Credits; paid plans keep completions unlimitedBuy additional credits or use a lower-cost model
GitHub Copilot Pro+$39/month$70 in total monthly GitHub AI CreditsBuy additional credits or change model
GitHub Copilot Max$100/month$200 in total monthly GitHub AI CreditsBuy additional credits or change model
Gemini Code Assist StandardAbout $22.80/month on monthly commitment; about $19/month on 12-month commitment1,500 combined agent-mode and Gemini CLI requests per user per dayWait for daily reset or request a quota adjustment where available
Gemini Code Assist EnterpriseAbout $54/month on monthly commitment; about $45/month on 12-month commitment2,000 combined agent-mode and Gemini CLI requests per user per dayWait for daily reset or request a quota adjustment where available
Cursor HobbyFreeLimited agent requests and limited Tab completionsUpgrade
Cursor Pro$20/month$20 of API-priced agent usage plus possible bonus usagePurchase usage at cost or upgrade
Cursor Pro+$60/month$70 of API-priced agent usage plus possible bonus usagePurchase usage at cost or upgrade
Cursor Ultra$200/month$400 of API-priced agent usage plus possible bonus usagePurchase additional usage at cost

The Claude individual prices and Max multipliers come from Anthropic’s plan guide. Anthropic also says Claude Code is covered by Pro and Max, while activity in Claude and Claude Code shares the same subscription limits; using an API key instead can create separate API charges (Claude Code billing guide).

OpenAI lists ChatGPT Plus at $20 on its ChatGPT pricing page, while its current Pro tiers guide lists Pro 5x at $100 and Pro 20x at $200. Both Pro tiers include Codex and the same core capabilities; their usage allowances differ. Codex’s current system is more important than the headline subscription price: most Plus and Pro accounts use token-based credits, with different rates for input, cached input, and output tokens (Codex rate card). OpenAI says included plan usage is consumed first; purchased credits can then extend Codex usage for eligible Plus and Pro accounts (flexible usage credits).

GitHub publishes individual prices and monthly AI-credit pools on its Copilot plans page. GitHub defines one AI credit as $0.01 and notes that chat, agents, code review, CLI, and other AI features draw from credits, while paid-plan code completions do not (Copilot billing documentation).

Google lists Gemini Code Assist Standard at $0.031232877 per license-hour on a monthly commitment and $0.026027397 on a 12-month commitment; Enterprise is $0.073972603 and $0.061643836 respectively. The monthly figures in the table are those official hourly rates multiplied by 730 hours (Google Cloud pricing). Google’s current quota page lists 1,500 combined agent-mode and Gemini CLI requests per user per day for Standard and 2,000 for Enterprise (quotas and limits). Consumer access through Gemini Code Assist for individuals was deprecated in June 2026, so it is not presented here as a current free coding-agent tier (license documentation).

Cursor’s pricing documentation states that agent usage is charged against the plan at model API prices: Pro includes $20, Pro+ $70, and Ultra $400 of agent usage, plus any bonus usage Cursor provides. Its public pricing page should be checked at purchase because displayed tiers and annual discounts can change.

Why the same monthly price does not mean the same capacity

Relative limits are not token guarantees

Claude’s 5x and 20x labels are relative to Pro capacity per session, not a promise of a fixed number of prompts. Anthropic says message consumption varies with prompt length, attached files, conversation length, model, and feature. It also applies weekly limits and may apply other caps (Max plan details). A short question and a repository-wide refactor should not be treated as equal units.

Credit pools are clearer, but model choice changes burn rate

GitHub and Cursor expose dollar-like pools, yet one request can consume a different amount depending on the model and work performed. GitHub says the cost of an interaction depends on model and complexity. Cursor similarly deducts model inference at API prices. A $70 pool is therefore a budget, not a fixed request count.

OpenAI’s Codex rate card makes the same point at token level. As of 2026-07-20, the official card lists separate credit rates per million tokens for GPT-5.6 Sol, Terra, and Luna, and cached input is cheaper than uncached input. Long context and verbose output can therefore change the number of tasks a subscription supports even when the seat price is unchanged.

Background and cloud agents may have separate costs

Do not assume every asynchronous feature is covered by the editor subscription. Cursor states that Background Agents are charged at the selected model’s API pricing and asks users to set a spend limit. GitHub agent, CLI, and review work all draw from AI credits. Claude can switch from subscription capacity to API or usage-credit billing only through a separate choice. These distinctions matter more for automated queues than for occasional interactive edits.

Which plan fits which usage pattern?

Occasional individual use

Start with GitHub Copilot Free or Cursor Hobby if the goal is to learn the workflow without a recurring charge. Free tiers are intentionally limited, so they are poor baselines for estimating sustained autonomous work. GitHub’s free plan is unusually explicit about 2,000 monthly completions and limited chat/agent access.

Daily coding with one primary agent

The $10–$20 range is the practical entry point: Copilot Pro, Cursor Pro, Claude Pro, or ChatGPT Plus. Choose based on the surrounding workflow, not a synthetic request count. Copilot is attractive when GitHub, IDE completion, CLI, and code review should share one product. Cursor is editor-centered and exposes an API-priced agent pool. Claude Pro and ChatGPT Plus also include broader non-coding assistants, but their coding limits are shared with other eligible usage.

Sustained parallel or repository-wide work

Heavy agent use can make $100–$200 plans more predictable, but only after measuring a real workload. Claude Max raises relative capacity; Copilot Max raises the included AI-credit pool; Cursor Ultra includes a larger API-priced pool; ChatGPT Pro 5x and 20x raise access while Codex still meters actual work. None should be described as unlimited autonomous coding.

Teams

Team plans add administration, policy, privacy, pooled billing, or indemnity rather than merely more prompts. This article focuses on individual plans because enterprise quotes and controls are not directly comparable. For a team purchase, run a two-week pilot and record cost per accepted change, not cost per generated answer.

A purchasing method that avoids false economy

  1. Collect 20 representative tasks: small fixes, tests, reviews, and two larger changes.
  2. Record input size, output size, retries, agent minutes, and whether the final patch passed tests.
  3. Calculate cost per accepted task and cost per merged change.
  4. Separate interactive editing from background, review, and parallel-agent usage.
  5. Set hard overage budgets before enabling automatic credit purchases.
  6. Repeat the sample when a vendor changes models or metering.

The cheapest plan is the one that produces verified changes at the lowest controlled cost. A low seat price with frequent retries can be more expensive than a higher tier, while a high-capacity plan is wasteful if most work is autocomplete. Treat the table as a shortlist, then benchmark with your own repositories and acceptance gates.

Bottom line

As of 2026-07-20, GitHub Copilot offers the lowest paid individual entry price in this group. Claude and ChatGPT bundle coding agents with broader assistants. Cursor gives a relatively transparent model-priced agent allowance inside an editor-first product. Google positions current Gemini Code Assist licenses for organizations. Heavy-use individual tiers cluster around $100–$200, but their units are fundamentally different.

Before subscribing, verify the official price, determine whether limits are relative or credit-based, check how overages are activated, and measure accepted output rather than raw prompts. Prices and specifications can change after 2026-07-20.